Hi rooters,
I have another question regarding minuit errors, I am performing a simultaneous fit to some generated data and for two slightly different data sets I get drastically different errors for some of my parameters. The output from minuit is as follows:
Data set 1
[code] FLOATING-POINT NUMBERS ASSUMED ACCURATE TO 1e-10
EIGENVALUES OF SECOND-DERIVATIVE MATRIX:
-5.3039e-03 1.1625e-01 1.4414e-01 1.8431e-01 3.6386e-01 4.1798e-01 9.4951e-01 9.5027e-01 9.9812e-01 1.0001e+00 1.0083e+00 1.0553e+00 1.9299e+00 2.4414e+00 3.4457e+00
MINUIT WARNING IN HESSE
============== MATRIX FORCED POS-DEF BY ADDING 0.008750 TO DIAGONAL.
Minimizer is Minuit / Migrad
Chi2 = 3.88031
NDf = 1785
Edm = 1.43516e-09
NCalls = 927
Par_0 = 0.17571 +/- 0.192738
Par_1 = 0.174079 +/- 0.19416
Par_2 = 0.168692 +/- 0.192903
Par_3 = 0.0726579 +/- 0.196776
Par_4 = 0.969488 +/- 1.48442
Par_5 = 0.965794 +/- 1.47849
Par_6 = 0.100463 +/- 0.323743
Par_7 = 0.0679001 +/- 0.299028
Par_8 = 0.0840538 +/- 0.277424
Par_9 = 0.969264 +/- 1.48433
Par_10 = 0.970621 +/- 1.48546
Par_11 = 0.050598 +/- 0.0861003
Par_12 = 0.399686 +/- 0.00859826
Par_13 = 0.051675 +/- 0.101636
Par_14 = 0.0528252 +/- 0.139513 [/code]
Data set 2:
[code] FLOATING-POINT NUMBERS ASSUMED ACCURATE TO 1e-10
Minimizer is Minuit / Migrad
Chi2 = 3.89375
NDf = 1785
Edm = 2.97045e-10
NCalls = 757
Par_0 = 0.171082 +/- 0.200302
Par_1 = 0.170135 +/- 0.199124
Par_2 = 0.163083 +/- 0.1978
Par_3 = 0.0773044 +/- 0.201293
Par_4 = 0.967315 +/- 0.177045
Par_5 = 0.972658 +/- 0.184761
Par_6 = 0.0873108 +/- 0.285436
Par_7 = 0.067153 +/- 0.281987
Par_8 = 0.092558 +/- 0.243022
Par_9 = 0.97447 +/- 0.182999
Par_10 = 0.963754 +/- 0.183037
Par_11 = 0.0477637 +/- 0.0869735
Par_12 = 0.399626 +/- 0.00880988
Par_13 = 0.0453092 +/- 0.104486
Par_14 = 0.0560008 +/- 0.141438 [/code]
This seems to happen whenever minuit says something like “MATRIX FORCED POS-DEF BY ADDING 0.008750 TO DIAGONAL.” Is there a way to get around this issue? or is there a reason why this happens?
One reason I suspect why this is happening is because there are elements in the correlation matrix which are near 1 (~0.97). The parameters however should be independent so is there any reason why minuit may believe that some parameters may be correlated when they are not. any insight would be much appreciated
cheers
Mark